Project Aim

Back out trends in the prevalence of Rifampicin resistant TB (RR-TB) over time using Xpert testing data in Sinan
Develop a relatively simple model that accounts for variation in roll-out of Xpert over time and across geographies (e.g. by state), as well as potential bias in who is getting tested with Xpert

Helpful context: Trends in Xpert testing overtime

As of 2019:
- ~ 40% of cases were being tested with Xpert, compared to ~ 3% in 2014 and 23.5% in 2015, on average
- Mato Grosso, Pará, and Pernambuco had the lowest average testing coverage
- Roraima, Amazonas, and Rondônia had the highest average testing coverage

State-level hierarchical generalized additive model (GAM) that models the prevalence of RR-TB positive cases per quarter among incident TB cases between 2014-2019
Fit smoothing functions to reduce the noise we were seeing in previous models
Models risk of positivity by characteristics of patient and municipality where they reside
- Note: Between 2014-2019, ~3,300 cases diagnosed outside of patient’s state of residence; ~88,000 cases diagnosed outside patient’s municipality of residence
Separate models for new TB cases, re-entry cases, and relapsed cases

result ~ s(state, bs = "re") + s(time) + s(time, by = state, id = 1) + age_cat + 
  hiv_status + sex + health_unit + bf_cat + urban_cat + has_prison

Random intercept for each state (patient state of residence)
A different smooth function for time by state with a shared smoothing parameter
Each state-level smoothing parameter varies around a grand smooth function for time to allow for pooling across states
Fixed effects for patient-level characteristics:
- Age
- HIV status
- Sex
- Level of health unit of diagnosis - Based on CNES merge
Fixed effects for municipality-level characteristics:
- Urbanicity (cat) - Percent of the population in urban setting (2010 census)
- Bolsa Familia coverage (cat) - Percent of the population benefiting from BF (BF: SAGICAD, 2018 - earliest year available; Denominator: 2010 Census)
- Presence of prison during year (SISPEN)
- FHS Coverage - Number of health teams per 4,000 people (Forthcoming)

Model 1 - Adjusted:
- Patient-level: HIV status, sex, age, health unit level
- Municipality: BF coverage, urbanicity, presence of prison
Model 2 - Adjusted; Restricted > 2015
Run separately by case type (e.g. new, relapse, re-entry) and for all cases

Model A - Adjusted:
- Patient-level: HIV status, sex, age (cat), health unit
- Municipality: Pct average BF coverage (cat), pct pop in urban area, presence of prison
Model B - Model A + interaction term (HIV, sex, age)
Model C - Model A, restricted to >2015

Note: The following figures show the same model output, only the Y axis changes to show variation within each state.

Note: The following figures show the same model output, only the Y axis changes to show variation within each state.

Add in FHS coverage
Try aggregating data to health region (to overcome noise when working with small municipalities)